All Downloads are FREE. Search and download functionalities are using the official Maven repository.

com.tangosol.net.partition.SimpleAssignmentStrategy Maven / Gradle / Ivy

There is a newer version: 24.09
Show newest version
/*
 * Copyright (c) 2000, 2021, Oracle and/or its affiliates.
 *
 * Licensed under the Universal Permissive License v 1.0 as shown at
 * http://oss.oracle.com/licenses/upl.
 */

package com.tangosol.net.partition;

import com.oracle.coherence.common.base.Logger;

import com.tangosol.coherence.config.Config;

import com.tangosol.net.Member;
import com.tangosol.net.PartitionedService;
import com.tangosol.net.ServiceInfo;

import com.tangosol.net.management.AnnotatedStandardEmitterMBean;
import com.tangosol.net.management.Registry;

import com.tangosol.util.Base;
import com.tangosol.util.ClassHelper;
import com.tangosol.util.Filter;
import com.tangosol.util.ImmutableArrayList;
import com.tangosol.util.ServiceEvent;
import com.tangosol.util.ServiceListener;
import com.tangosol.util.SubSet;
import com.tangosol.util.SynchronousListener;

import com.tangosol.util.comparator.ChainedComparator;
import com.tangosol.util.comparator.InverseComparator;

import com.tangosol.util.filter.AllFilter;
import com.tangosol.util.filter.AlwaysFilter;
import com.tangosol.util.filter.NotFilter;

import java.util.ArrayList;
import java.util.Arrays;
import java.util.Collections;
import java.util.Comparator;
import java.util.Date;
import java.util.HashMap;
import java.util.HashSet;
import java.util.Iterator;
import java.util.List;
import java.util.Map;
import java.util.Map.Entry;
import java.util.Set;
import java.util.SortedMap;
import java.util.TreeMap;

import javax.management.NotCompliantMBeanException;
import javax.management.Notification;


/**
 * SimpleAssignmentStrategy is a PartitionAssignmentStrategy that attempts to
 * balance the partition distribution based on the number of primary and backup
 * partitions owned.  The SimpleAssignmentStrategy will attempt to ensure
 * machine-safety, but only if a balanced "safe" distribution is achievable.
 * 

* The SimpleAssignmentStrategy is an extensible implementation of the internal * distribution algorithm that was available prior to Coherence 3.7. * * @author rhl 2010.11.08 * @since Coherence 3.7 */ public class SimpleAssignmentStrategy implements PartitionAssignmentStrategy, SimpleStrategyMBean { // ----- constructors ------------------------------------------------- /** * Default constructor. */ public SimpleAssignmentStrategy() { } // ----- accessors ---------------------------------------------------- /** * Return the DistributionManager. * * @return the DistributionManager */ public DistributionManager getManager() { return m_manager; } /** * Return the last AnalysisContext. * * @return the last AnalysisContext */ public AnalysisContext getLastAnalysisContext() { return m_ctxLast; } /** * Set the last AnalysisContext. * * @param ctx the AnalysisContext */ protected void setLastAnalysisContext(AnalysisContext ctx) { m_ctxLast = ctx; } /** * Return the set of ownership-enabled members present when the analysis was * last considered. * * @return the set of ownership-enabled members when analysis was last * considered */ public Set getLastOwnershipMembers() { return m_setOwnersLast; } /** * Record the set of ownership-enabled members present when the analysis was * last considered. * * @param setOwners the set of ownership-enabled members */ protected void setLastOwnershipMembers(Set setOwners) { m_setOwnersLast = setOwners; } /** * Return the amount of time in ms to delay the analysis after a member has * joined. This delay could be used to "dampen" the reactivity of the * strategy to membership changes. * * @return the amount of time in ms to delay the analysis after a member joins */ protected long getMemberJoinDelay() { return 1000L; } /** * Return the amount of time in ms to delay the analysis. * * @return the amount of time in ms to delay the analysis */ protected long getSuggestionDelay() { return 60000L; } /** * Return the amount of time in ms to delay the analysis after a * distribution suggestion has been made and before it is carried out. This * delay could be used to "dampen" the volatility of the strategy by * allowing sufficient time for in-flight transfers to complete prior to * considering further recommendations. * * @return the amount of time in ms to delay the analysis after a suggestion * is made */ protected long getSuggestionCompletionDelay() { return m_cPlanCompletionDelay; } // ----- PartitionAssignmentStrategy methods -------------------------- /** * {@inheritDoc} */ public void init(DistributionManager manager) { class ServiceStoppedListener implements ServiceListener, SynchronousListener { public void serviceStarting(ServiceEvent evt) {} public void serviceStarted(ServiceEvent evt) {} public void serviceStopping(ServiceEvent evt) {} public void serviceStopped(ServiceEvent evt) { unregisterMBean(); } } m_manager = manager; m_cPlanCompletionDelay = getPartitionCount() < 1 << 14 ? 60_000L : 300_000L; registerMBean(); manager.getService().addServiceListener(new ServiceStoppedListener()); } /** * {@inheritDoc} */ public void analyzeOrphans(final Map mapConstraints) { AnalysisContext ctx = instantiateAnalysisContext(); int cPartitions = getPartitionCount(); PartitionSet partsLost = new PartitionSet(cPartitions); Member[] aMember = ctx.getOwnershipMembersList(); for (int iPart = 0; iPart < cPartitions; iPart++) { Ownership owners = ctx.getPartitionOwnership(iPart); int nOwner = owners.getPrimaryOwner(); if (nOwner == 0) { Comparator comparator = chainComparators(ctx.instantiateStrengthComparator(owners), ctx.instantiateLoadComparator(true), ctx.instantiateDefaultComparator()); // this partition is orphaned; assemble the set of members which // it could be assigned to ordered by strength and load int iPartition = iPart; int cMembers = filterSort(aMember, comparator, member -> { PartitionSet parts = mapConstraints.get(member); return parts != null && parts.contains(iPartition); }); if (cMembers == 0) { // nothing to recover from; simply balance the assignments partsLost.add(iPart); cMembers = filterSort(aMember, comparator, AlwaysFilter.INSTANCE); } if (cMembers > 0) { ctx.transitionPartition(iPart, 0, null, aMember[0]); } } } ctx.suggestDistribution(); if (!partsLost.isEmpty()) { emitLossNotification(partsLost); // retain knowledge of partitions that were orphaned to prioritize // their distribution when aiming to achieve balance thus minimizing // transfer cost ctx.setOrphanedPartitions(partsLost); } // force an analysis of the ownership immediately following a partition // recovery or loss; we almost certainly need to move something around ctx.setAnalysisDelay(0L); setLastAnalysisContext(ctx); } /** * {@inheritDoc} */ public long analyzeDistribution() { AnalysisContext ctx = instantiateAnalysisContext(); ctx.copyTransients(getLastAnalysisContext()); long cDelay = ctx.calculateAnalysisDelay(); if (cDelay <= 0L) { cDelay = analyzeDistribution(ctx); ctx.resetTransients(); setLastAnalysisContext(ctx); } // record the ownership member-set setLastOwnershipMembers(ctx.getOwnershipMembers()); m_fRefresh = true; return cDelay; } /** * Analyze the distribution and suggest the resulting distribution * to distribution manager. * * @param ctx the analysis context * * @return the time interval before the next desired analysis, or -1 */ protected long analyzeDistribution(AnalysisContext ctx) { // do a initial two server distribution to reduce the backup "fan out" primeDistribution(ctx); long cSuggestDelay = analyze(ctx); ctx.suggestDistribution(); ctx.setCompletedTime(Base.getSafeTimeMillis()); return cSuggestDelay; } /** * Optimize the distribution with minimum backup fan-out by distributing * the partitions among first two strong members. * * @param ctx the analysis context */ protected void primeDistribution(AnalysisContext ctx) { // If we delayed the initial (after the coordinator assigned all the // partitions to itself) distribution and have more than two new members // that have joined, the effect of an immediate call to "analyze" would // be a large "fan-out" of the backups. It is a result of the current // distribution algorithm and needs to be improved, but the scope ot // that project is rather large. // // In turn, the high backup fan-out results in a large number of backup // messages produced by any bulk operations that span multiple partitions // (see COH-14955). // // However, if we start with just two members that evenly divide the // partitions among themselves, the "fan-out" effect of the distribution // algorithm is dramatically lower. Hence our work around the fan-out // problem is to call the distribution analysis twice: first by pretending // that only two members exist and then proceed to the actual membership. Member memberCoordinator = getManager().getService().getCluster().getLocalMember(); // first, quickly check that the current distribution is in the initial state - // the coordinator owns everything and no backups are owned if (!ctx.isInitialDistribution(memberCoordinator)) { return; } if (ctx.getOwnershipMembers().size() <= 2 || ctx.getActualBackupCount() != 1) { return; } Member memberStrong = null; for (Member member : ctx.getOwnershipMembers()) { if (ctx.isStrong(memberCoordinator, member)) { memberStrong = member; break; } } // we must find a "strong" second, because the strength is based on // the context itself assert(memberStrong != null); // prime the distribution with a "two servers" scenario ctx.primeDistribution(memberCoordinator, memberStrong); analyze(ctx); // revert the context info to the original state ctx.initialize(); } /** * Analyze the distribution using the specified analysis context. * * @param ctx the analysis context * * @return the time interval before the next desired analysis, or -1 */ protected long analyze(AnalysisContext ctx) { long cSuggestDelay = getSuggestionDelay(); checkLeaving(ctx); validateBackups(ctx); // as of Coherence 12.2.1.1, the two-server case with backup is treated // differently (SE One contract) if (m_fTrivialDistribution && checkSimpleDistribution(ctx)) { return cSuggestDelay; } checkPrimaryBalance(ctx); BackupStrength strengthOrig = null; if (ctx.getActualBackupCount() > 0) { checkEndangered(ctx); int cChanges; int cIters = 0; int cIterMax = 10; int cVariancePrev = getVariance(ctx, false); do { cChanges = 0; // redistribute until backup strength & balance are stable cChanges += checkBackupStrong (ctx); cChanges += checkBackupBalance(ctx); if (cChanges == 0) { // distribution may still not balanced; this should happen very rarely cChanges = checkBackupOverloaded(ctx); } if (cChanges == 0) { // reached fixed-point break; } // calculate the variance and make sure that we still making // progress toward the fixed-point int cVarianceCur = getVariance(ctx, false); if (cIters++ > cIterMax && cVarianceCur >= cVariancePrev) { if (strengthOrig == null) { // after cIterMax iterations, if we are not monotonically // proceeding towards the fixed-point, slightly disturb the distribution strengthOrig = ctx.getBackupStrength(); checkBackupOverloaded(ctx); } else { // after 10 iterations after the reshuffle, if we // still couldn't reach the balance point, give up and decrease // the strength as a fail-safe to prevent an infinite loop; // reschedule the next analysis relatively soon and // log a soft-assert, as this shouldn't really happen Logger.err("Failed to find a partition assignment to satisfy " + strengthOrig + " among the member-set " + ctx.getOwnershipMembers() + "; weakening the backup-strength"); ctx.setBackupStrength(strengthOrig.getWeaker()); strengthOrig = null; cSuggestDelay = 1000L; } cIters = 0; } cVariancePrev = cVarianceCur; } while (true); } return cSuggestDelay; } /** * Analyze the distribution for the special two-server case using the * specified analysis context. * * @param ctx the analysis context * * @return true if the simple distribution is in effect */ protected boolean checkSimpleDistribution(AnalysisContext ctx) { DistributionManager manager = getManager(); if (manager.getOwnershipMembers().size() > 2 || getBackupCount() == 0) { // more than two servers, or no backup is not SE One return m_fTrivialDistribution = false; } if (!ctx.getLeavingOwners().isEmpty() || ctx.getOwnershipMembers().size() == 1) { // defer to the standard algorithm, but allow to come back return false; } Member member1 = manager.getService().getOwnershipSenior(); Member member2 = null; for (Member member : ctx.getOwnershipMembers()) { if (member != member1) { member2 = member; break; } } if (!ctx.getOwnedPartitions(member2, 0).isEmpty()) { // we should never come here, but just in case... return m_fTrivialDistribution = false; } PartitionSet partsBackup2 = ctx.getOwnedPartitions(member2, 1); if (!partsBackup2.isFull()) { // calculate the partitions that the backup member is not a backup owner yet // and suggest a backup transfer PartitionSet parts = new PartitionSet(partsBackup2); parts.invert(); for (int iPart = parts.next(0); iPart >= 0; iPart = parts.next(iPart + 1)) { ctx.transitionPartition(iPart, 1, member1, member2); } } return true; } /** * {@inheritDoc} */ public String getDescription() { AnalysisContext ctxLast = getLastAnalysisContext(); StringBuilder sb = new StringBuilder(); if (ctxLast != null) { sb.append("Fair-Share=").append(ctxLast.getFairShare(true)) .append("(primary) ").append(ctxLast.getFairShare(false)) .append("(backup)").append(", ") .append("Target Backup-Strength=") .append(ctxLast.getBackupStrength().getDescription()); } return sb.toString(); } // ----- Object methods ----------------------------------------------- /** * {@inheritDoc} */ public String toString() { return ClassHelper.getSimpleName(getClass()) + '{' + getDescription() + '}'; } // ----- internal ----------------------------------------------------- /** * Check for any service members that are leaving, and adjust the * distribution plan accordingly. *

* Partitions owned by leaving members must be transferred to other members * before the departing members are able to shutdown. * * @param ctx the AnalysisContext */ protected void checkLeaving(AnalysisContext ctx) { Set setLeaving = ctx.getLeavingOwners(); Member[] aOwners = ctx.getOwnershipMembersList(); int cBackupsConfig = getBackupCount(); if (setLeaving.isEmpty()) { // no more departing (awaiting graceful exit pending transfer) members return; } // iterate the leaving members to transfer away the owned partitions for (Member memberLeaving : (Set) setLeaving) { // 1. transfer away the owned primary partitions PartitionSet partsPrime = ctx.getOwnedPartitions(memberLeaving, 0); partition_loop: for (int iPart = partsPrime.next(0); iPart >= 0; iPart = partsPrime.next(iPart + 1)) { Ownership owners = ctx.getPartitionOwnership(iPart); // There are 2 cases: either there is some other backup owner, // or the partition is endangered. If there is a backup owner, // promote it to primary. Otherwise, pick the lightest member // to force the transfer to. for (int iStore = cBackupsConfig; iStore >= 1; iStore--) { int nOwner = owners.getOwner(iStore); if (nOwner != 0) { Member memberTo = getMember(nOwner); // make the previous backup member the new primary owner ctx.transitionPartition(iPart, 0, memberLeaving, memberTo); // endanger the storage-index previously owned ctx.transitionPartition(iPart, iStore, memberTo, null); continue partition_loop; } } // the partition had zero backups; force the transfer to the // most lightly loaded member Arrays.sort(aOwners, ctx.instantiateLoadComparator(true)); ctx.transitionPartition(iPart, 0, memberLeaving, aOwners[0]); } // 2. set the backup partitions to "endangered" so that they will be // handled by the next pass for (int iStore = 1; iStore <= cBackupsConfig; iStore++) { PartitionSet parts = ctx.getOwnedPartitions(memberLeaving, iStore); for (int iPart = parts.next(0); iPart >= 0; iPart = parts.next(iPart + 1)) { // set the previously owned storage-index to endangered ctx.transitionPartition(iPart, iStore, memberLeaving, null); } } } } /** * Check if there are enough ownership members to maintain the configured * backup count, reducing the number of backups otherwise. * * @param ctx the AnalysisContext */ protected void validateBackups(AnalysisContext ctx) { int cPartitions = getPartitionCount(); int cBackupsActual = ctx.getActualBackupCount(); int cBackupsConfig = getBackupCount(); // if departing (or departed) members leave us without enough members to // maintain full backups, prune the backup storage. // // Note: this logic emulates the automatic "pruning" performed by the // PartitionedService. See PartitionedService.validateBackupCount() if (cBackupsConfig != cBackupsActual) { for (int iPart = 0; iPart < cPartitions; iPart++) { Ownership owners = ctx.getPartitionOwnership(iPart); for (int iStore = 1, iStoreValid = 1; iStore <= cBackupsConfig; iStore++) { int nBackupOwner = owners.getOwner(iStore); if (nBackupOwner != 0) { if (iStore > iStoreValid) { Member memberThis = getMember(nBackupOwner); // move this backup index down from iStore to iStoreValid ctx.transitionPartition(iPart, iStoreValid, null, memberThis); // set the iStore storage index to endangered ctx.transitionPartition(iPart, iStore, memberThis, null); } iStoreValid++; } } } } } /** * Check the distribution to ensure that primary the partition load is * balanced. * * @param ctx the analysis context */ protected void checkPrimaryBalance(AnalysisContext ctx) { Member[] aMembersOverload = ctx.getOwnershipMembersList(); Member[] aMembersTarget = aMembersOverload.clone(); // repeat this until there are no more primary distributions to perform. // This will ensure a "refined" distribution. int cChanges; do { cChanges = 0; // sort the member list in decreasing order of primary partition load int cOverloaded = filterSort( aMembersOverload, new InverseComparator(ctx.instantiateLoadComparator(true)), ctx.instantiateOverloadedFilter(true)); for (int i = 0; i < cOverloaded; i++) { Member memberFrom = aMembersOverload[i]; PartitionSet partsAll = new PartitionSet(ctx.getOwnedPartitions(memberFrom, 0)); // try orphaned partitions first PartitionSet partsOrphaned = ctx.collectOrphaned(partsAll); partsAll.remove(partsOrphaned); cChanges += doBalancePrimary( ctx, memberFrom, partsOrphaned, aMembersTarget); // try endangered partitions first PartitionSet partsEndangered = ctx.collectEndangered(partsAll); partsAll.remove(partsEndangered); cChanges += doBalancePrimary( ctx, memberFrom, partsEndangered, aMembersTarget); // next try vulnerable partitions PartitionSet partsWeak = ctx.collectWeak(partsAll); partsAll.remove(partsWeak); cChanges += doBalancePrimary( ctx, memberFrom, partsWeak, aMembersTarget); // lastly, any partition cChanges += doBalancePrimary( ctx, memberFrom, partsAll, aMembersTarget); } } while (cChanges > 0); } /** * Do balancing transfers for primary distribution. * * @param ctx the analysis context * @param memberFrom the member to transfer partitions from * @param parts the set of partitions from which to transfer * @param aMembersTarget the (unordered) array of members * * @return the number of changes (transfers) that were made */ protected int doBalancePrimary(AnalysisContext ctx, Member memberFrom, PartitionSet parts, Member[] aMembersTarget) { int cFairShare = ctx.getFairShare(true); int cLoadMemberFrom = ctx.getMemberLoad(memberFrom, true); int cChanges = 0; for (int iPart = parts.next(0); iPart >= 0; iPart = parts.next(iPart + 1)) { if (cLoadMemberFrom < cFairShare) { break; } // consider partitions to transfer Ownership owners = (Ownership) ctx.getPartitionOwnership(iPart).clone(); int cLoad = ctx.getPartitionLoad(iPart, true); // clear the primary owner before evaluating potential replacements owners.setOwner(0, 0); // select the list of underloaded members, sorted by decreasing // strength (strongest-first) int cUnderloaded; try { cUnderloaded = filterSort( aMembersTarget, chainComparators( ctx.instantiateStrengthComparator(owners), ctx.instantiateLoadComparator(true), ctx.instantiateDefaultComparator()), ctx.instantiateUnderloadedFilter(true)); } catch (Throwable t) { StringBuilder sb = new StringBuilder("Member array: ["); for (Member member : aMembersTarget) { sb.append(member == null ? "null" : member.getId()).append(','); } sb.replace(sb.length() - 1, sb.length(), "]") .append("\nPartition Id: ").append(iPart); Logger.err(sb.toString()); throw Base.ensureRuntimeException(t); } // find a new primary owner for this partition for (int i = 0; i < cUnderloaded; i++) { Member memberTo = aMembersTarget[i]; int cLoadMemberTo = ctx.getMemberLoad(memberTo, true); // only if it balances load if (cLoadMemberTo + cLoad < cLoadMemberFrom) { ctx.transitionPartition(iPart, 0, memberFrom, memberTo); cLoadMemberFrom -= cLoad; ++cChanges; break; } } } return cChanges; } /** * Check the distribution to ensure that backups are created for any * "endangered" partitions. *

* A partition is "endangered" if it is incompletely backed up (e.g. some * backup copies do not exist). * * @param ctx the analysis context */ protected void checkEndangered(AnalysisContext ctx) { int cBackups = ctx.getActualBackupCount(); int cPartitions = getPartitionCount(); Member[] aMember = ctx.getOwnershipMembersList(); for (int iPart = 0; iPart < cPartitions; iPart++) { Ownership owners = ctx.getPartitionOwnership(iPart); Member memberPrimary = getMember(owners.getPrimaryOwner()); Base.azzert(memberPrimary != null); // there shouldn't be any orphans for (int iStore = 1; iStore <= cBackups; iStore ++) { if (owners.getOwner(iStore) != 0) { continue; } // sort the member array and select the safe members, ordered by // strength and load int cSafe = filterSort( aMember, chainComparators(ctx.instantiateStrengthComparator(owners), ctx.instantiateLoadComparator(false), ctx.instantiateDefaultComparator()), ctx.instantiateNotOwnedFilter(owners)); Base.azzert(cSafe > 0, "Failed to find a member to receive backup(" + iStore + ") transfer of endangered partition " + iPart + ", " + owners); ctx.transitionPartition(iPart, iStore, null, aMember[0]); } } } /** * Check that the backups are strong. * * @param ctx the analysis context * * @return the number of changes (transfers) that were made */ protected int checkBackupStrong(AnalysisContext ctx) { int cPartitions = getPartitionCount(); int cBackups = ctx.getActualBackupCount(); int cChanges = 0; Member[] aMembersTarget = ctx.getOwnershipMembersList(); for (int iPart = 0; iPart < cPartitions; iPart++) { for (int iStore = 1; !ctx.isPartitionStrong(iPart) && iStore <= cBackups; iStore++) { // try to find a new strong backup owner for the specified partition Ownership owners = (Ownership) ctx.getPartitionOwnership(iPart).clone(); Member memberFrom = getMember(owners.getOwner(iStore)); // clear the backup owner before evaluating potential replacements owners.setOwner(iStore, 0); // pre-filter the member array for safety and determine if full // "safety" is achievable int cCandidate = filterArray(aMembersTarget, ctx.instantiateSafetyFilter(owners, iStore)); if (cCandidate == 0) { // No member will produce a fully safe configuration after // a single transfer. This is possible if, for example, the // topology is: // Machine1: 1, 2, 3 // Machine2: 4, 5, 6 // Machine3: 7, 8, 9 // // If a partition starts with ownership (4,5,6), no single // transfer will render it "safe". cCandidate = aMembersTarget.length; } // first-pass: find the "strongest" safe, underloaded member int cUnderloaded = filterSort( aMembersTarget, cCandidate, chainComparators(ctx.instantiateStrengthComparator(owners), ctx.instantiateLoadComparator(false), ctx.instantiateDefaultComparator()), new AllFilter(new Filter[] { // Note: array is pre-filtered for safe members ctx.instantiateUnderloadedFilter(false), ctx.instantiateNotOwnedFilter(owners) })); if (cUnderloaded > 0) { ctx.transitionPartition(iPart, iStore, memberFrom, aMembersTarget[0]); ++cChanges; continue; } // second-pass: no strong underloaded members; find the least // overloaded safe member int cOverloaded = filterSort( aMembersTarget, cCandidate, chainComparators(ctx.instantiateLoadComparator(false), ctx.instantiateStrengthComparator(owners), ctx.instantiateDefaultComparator()), new AllFilter(new Filter[] { // Note: array is pre-filtered for safe members ctx.instantiateOverloadedFilter(false), ctx.instantiateNotOwnedFilter(owners) })); if (cOverloaded > 0) { ctx.transitionPartition(iPart, iStore, memberFrom, aMembersTarget[0]); ++cChanges; continue; } } } return cChanges; } /** * Check that the distribution of backup partitions is balanced. * * @param ctx the analysis context * * @return the number of changes (transfers) that were made */ protected int checkBackupBalance(final AnalysisContext ctx) { int cBackups = ctx.getActualBackupCount(); Member[] aMembersOverload = ctx.getOwnershipMembersList(); Member[] aMembersTarget = aMembersOverload.clone(); int cFairShare = ctx.getFairShare(false); int cChanges = 0; // sort the overloaded members in decreasing backup load order int cOverloaded = filterSort( aMembersOverload, new InverseComparator(ctx.instantiateLoadComparator(false)), ctx.instantiateOverloadedFilter(false)); member_loop: for (int i = 0; i < cOverloaded; i++) { Member memberFrom = aMembersOverload[i]; int cLoadMemberFrom = ctx.getMemberLoad(memberFrom, false); for (int iStore = 1; iStore <= cBackups; iStore++) { PartitionSet parts = ctx.getOwnedPartitions(memberFrom, iStore); for (int iPart = parts.next(0); iPart >= 0; iPart = parts.next(iPart + 1)) { int cLoad = ctx.getPartitionLoad(iPart, false); Ownership owners = (Ownership) ctx.getPartitionOwnership(iPart).clone(); // clear the backup owner before evaluating potential replacements owners.setOwner(iStore, 0); // select the underloaded and strong members, ordered by load int cUnderloaded = filterSort( aMembersTarget, chainComparators(ctx.instantiateLoadComparator(false), ctx.instantiateStrengthComparator(owners), ctx.instantiateDefaultComparator()), new AllFilter(new Filter[] { ctx.instantiateSafetyFilter(owners, iStore), ctx.instantiateUnderloadedFilter(false), ctx.instantiateNotOwnedFilter(owners) })); for (int j = 0; j < cUnderloaded; j++) { Member memberTo = aMembersTarget[j]; int cLoadMemberTo = ctx.getMemberLoad(memberTo, false); // only if it balances load if (cLoadMemberTo + cLoad < cLoadMemberFrom) { ctx.transitionPartition(iPart, iStore, memberFrom, memberTo); cLoadMemberFrom -= cLoad; ++cChanges; break; } } if (cLoadMemberFrom < cFairShare) { continue member_loop; } } } } return cChanges; } /** * Check if the distribution of backup partitions is balanced. If not, * disturb the distribution by moving a partition from the overloaded member * to another member that retains partition strength. * * @param ctx the analysis context * * @return the number of unbalanced partitions that need to be transferred */ protected int checkBackupOverloaded(final AnalysisContext ctx) { Member[] aMembers = (Member[]) Base.randomize(ctx.getOwnershipMembersList()); int cFairShareBackup = ctx.getFairShare(false); int cBackup = getBackupCount(); int cOverload = 0; Member memberOverloaded = null; for (int i = 0; i < aMembers.length; i++) { Member memberFrom = aMembers[i]; int cBackupLoad = ctx.getMemberLoad(memberFrom, false); int cLoad = cBackupLoad - cFairShareBackup; if (cLoad > 0) { cOverload += cLoad; memberOverloaded = memberFrom; break; } } if (memberOverloaded != null) { PartitionSet partsBackup = new PartitionSet(getPartitionCount()); for (int iStore = 1; iStore <= cBackup; iStore++) { partsBackup.add(ctx.getOwnedPartitions(memberOverloaded, iStore)); } member_loop: for (int i = 0; i < aMembers.length; i++) { Member memberTo = aMembers[i]; if (memberOverloaded != memberTo) { PartitionSet partOwned = ctx.getOwnedPartitions(memberTo, 0); for (int iStore = 1; iStore <= cBackup; iStore++) { if (!partOwned.intersects(partsBackup) && ctx.isStrong(memberOverloaded, memberTo)) { ctx.transitionPartition(partsBackup.next(0), iStore, memberOverloaded, memberTo); break member_loop; } } } } } return cOverload; } // ----- helpers ------------------------------------------------------ /** * Return the PartitionedService Member with the specified mini-id. * * @param nMemberId the mini-id * * @return the PartitionedService Member with the specified mini-id, or null */ protected Member getMember(int nMemberId) { return getManager().getMember(nMemberId); } /** * Return the maximum load variance in the partition assignments represented * by the analysis context. The maximum variance is the difference in load * between the 'lightest' and 'heaviest' members. * * @param ctx the analysis context * @param fPrimary true iff the "primary" load variance should be computed * * @return the maximum variance */ protected int getVariance(AnalysisContext ctx, boolean fPrimary) { // select the list of underloaded members, sorted by decreasing // strength (strongest-first) Member[] aMembers = ctx.getOwnershipMembersList(); int cMembers = aMembers.length; Arrays.sort(aMembers, 0, cMembers, chainComparators( ctx.instantiateLoadComparator(fPrimary), ctx.instantiateDefaultComparator())); return ctx.getMemberLoad(aMembers[cMembers - 1], fPrimary) - ctx.getMemberLoad(aMembers[0], fPrimary); } /** * Helper method to return a Comparator chaining the specified comparators. * * @param comp1 the first comparator * @param comp2 the second comparator * * @return a chained comparator */ protected static Comparator chainComparators(Comparator comp1, Comparator comp2) { return new ChainedComparator(comp1, comp2); } /** * Helper method to return a Comparator chaining the specified comparators. * * @param comp1 the first comparator * @param comp2 the second comparator * @param comp3 the third comparator * * @return a chained comparator */ protected static Comparator chainComparators( Comparator comp1, Comparator comp2, Comparator comp3) { return new ChainedComparator(comp1, comp2, comp3); } /** * Filter the elements in the specified array and sort any matching elements * using the specified comparator. All matching results will be compacted to * the front of the array. The order of results not matching the filter is * undefined. * * @param ao the object array to sort and filter * @param comparator the comparator to order the elements * @param filter the filter to use to filter the results * * @return the number of elements matching the specified filter */ protected static int filterSort(Object[] ao, Comparator comparator, Filter filter) { return filterSort(ao, ao.length, comparator, filter); } /** * Filter the specified array elements and sort any matching elements * using the specified comparator. All matching results will be compacted to * the front of the array. The order of results not matching the filter is * undefined. * * @param ao the object array to sort and filter * @param cElems the number of elements to filter and sort * @param comparator the comparator to order the elements * @param filter the filter to use to filter the results * * @return the number of elements matching the specified filter */ protected static int filterSort(Object[] ao, int cElems, Comparator comparator, Filter filter) { cElems = filterArray(ao, cElems, filter); if (cElems > 1) { Arrays.sort(ao, 0, cElems, comparator); } return cElems; } /** * Apply the specified filter to the elements of the specified array. All * matching results will be compacted to the front of the array in a "stable" * manner. The order of results not matching the filter may not be * preserved. * * @param ao the object array to apply the filter to * @param filter the filter to apply * * @return the number of elements matching the specified filter */ protected static int filterArray(Object[] ao, Filter filter) { return filterArray(ao, ao.length, filter); } /** * Apply the specified filter to the specified array elements. All * matching results will be compacted to the front of the array in a "stable" * manner. The order of results not matching the filter may not be * preserved. * * @param ao the object array to apply the filter to * @param cElems the number of elements to filter * @param filter the filter to apply * * @return the number of elements matching the specified filter */ protected static int filterArray(Object[] ao, int cElems, Filter filter) { int iShift = 0; for (int i = 0; i < cElems; i++) { if (!filter.evaluate(ao[i])) { ++iShift; } else if (iShift > 0) { Object oTemp = ao[i - iShift]; ao[i - iShift] = ao[i]; ao[i] = oTemp; } } return cElems - iShift; } // ----- SimpleStrategyMBean interface ---------------------------------- /** * {@inheritDoc} */ public int getPartitionCount() { return getManager().getService().getPartitionCount(); } /** * {@inheritDoc} */ public int getBackupCount() { return getManager().getService().getBackupCount(); } /** * {@inheritDoc} */ public int getServiceNodeCount() { return getManager().getService().getOwnershipEnabledMembers().size(); } /** * {@inheritDoc} */ public int getServiceMachineCount() { AnalysisContext ctx = getLastAnalysisContext(); return ctx == null ? 0 : ctx.getBackupStrength().getMachineCount(); } /** * {@inheritDoc} */ public int getServiceRackCount() { AnalysisContext ctx = getLastAnalysisContext(); return ctx == null ? 0 : ctx.getBackupStrength().getRackCount(); } /** * {@inheritDoc} */ public int getServiceSiteCount() { AnalysisContext ctx = getLastAnalysisContext(); return ctx == null ? 0 : ctx.getBackupStrength().getSiteCount(); } /** * {@inheritDoc} */ public String getHAStatus() { // TODO: move this functionality into a helper class try { return (String) ClassHelper.invoke(getManager().getService(), "getBackupStrengthName", ClassHelper.VOID_PARAMS); } catch (Exception e) { return ""; } } /** * {@inheritDoc} */ public int getHAStatusCode() { try { return (int) ClassHelper.invoke(getManager().getService(), "getBackupStrength", ClassHelper.VOID_PARAMS); } catch (Exception e) { return -1; } } /** * {@inheritDoc} */ public String getHATarget() { AnalysisContext ctx = getLastAnalysisContext(); return ctx == null ? MSG_NO_RESULT : ctx.getBackupStrength().getDescription(); } /** * {@inheritDoc} */ public int getFairShareBackup() { AnalysisContext ctx = getLastAnalysisContext(); return ctx == null ? 0 : ctx.getFairShare(false); } /** * {@inheritDoc} */ public int getFairSharePrimary() { AnalysisContext ctx = getLastAnalysisContext(); return ctx == null ? 0 : ctx.getFairShare(true); } /** * {@inheritDoc} */ public String getStrategyName() { return getClass().getSimpleName(); } /** * {@inheritDoc} */ public Date getLastAnalysisTime() { AnalysisContext ctx = getLastAnalysisContext(); return ctx == null ? new Date(0) : new Date(ctx.getCompletedTime()); } /** * {@inheritDoc} */ public int getCoordinatorId() { try { return getManager().getService().getCluster().getLocalMember().getId(); } catch (NullPointerException e) { return 0; } } /** * {@inheritDoc} */ public int getRemainingDistributionCount() { int cScheduled = 0; Map mapScheduled = collectScheduledDistributions(); for (Map.Entry entry : mapScheduled.entrySet()) { for (PartitionSet parts : entry.getValue()) { cScheduled += parts == null ? 0 : parts.cardinality(); } } return cScheduled; } /** * {@inheritDoc} */ public long getAveragePartitionSizeKB() { return updateCompositeStats().getAveragePartitionSize(); } /** * {@inheritDoc} */ public long getMaxPartitionSizeKB() { return updateCompositeStats().getMaxPartitionSize(); } /** * {@inheritDoc} */ public long getMaxStorageSizeKB() { return updateCompositeStats().getMaxStorageSize(); } /** * {@inheritDoc} */ public long getAverageStorageSizeKB() { return updateCompositeStats().getAverageStorageSize(); } /** * {@inheritDoc} */ public int getMaxLoadNodeId() { return updateCompositeStats().getMaxLoadNodeId(); } /** * Return a JMXPartitionStats that contains calculated MBean Attributes, updated * periodically. * * @return JMXPartitionStats with calculated MBean Attributes */ protected JMXPartitionStats updateCompositeStats() { JMXPartitionStats stats = m_statsPartition; if (stats == null || m_fRefresh) { if (stats == null) { m_statsPartition = stats = new JMXPartitionStats(); } stats.calculateJMXPartitionStats(); m_fRefresh = false; } return stats; } /** * {@inheritDoc} */ public String reportScheduledDistributions(boolean fVerbose) { Map mapScheduled = collectScheduledDistributions(); if (mapScheduled.isEmpty()) { return getLastAnalysisContext() == null ? MSG_NO_RESULT : MSG_NO_PENDING; } StringBuilder sb = new StringBuilder(); sb.append("Partition Distributions Scheduled for Service \"") .append(getManager().getService().getInfo().getServiceName()) .append("\"\n"); // 1. group members by machine Map> mapByMachine = new HashMap<>(); for (Member member : mapScheduled.keySet()) { String sMachine = member.getMachineName(); sMachine = sMachine == null ? Integer.toString(member.getMachineId()) : sMachine; List listMembers = mapByMachine.get(sMachine); if (listMembers == null) { mapByMachine.put(sMachine, listMembers = new ArrayList<>()); } listMembers.add(member); } // 2. for each machine, summarize the scheduled distributions per member for (Entry> entryMachine : mapByMachine.entrySet()) { sb.append("\nMachine ").append(entryMachine.getKey()); for (Member member : entryMachine.getValue()) { sb.append("\n Member ").append(member.getId()).append(":"); for (int iStore = 0, cBackups = getBackupCount(); iStore <= cBackups; iStore++) { PartitionSet partsScheduled = mapScheduled.get(member)[iStore]; if (partsScheduled != null) { Map mapOwners = splitByOwner(partsScheduled); int cScheduled = partsScheduled.cardinality(); if (cScheduled > 0) { String sStore = iStore == 0 ? " Primary" : " Backup" + (cBackups == 1 ? "" : ("[" + iStore + "]")); sb.append("\n - scheduled to receive ").append(cScheduled) .append(sStore).append(" partitions:"); for (Entry entrySource : mapOwners.entrySet()) { Integer IOwner = entrySource.getKey(); PartitionSet parts = entrySource.getValue(); sb.append("\n -- ").append(parts.cardinality()) .append(" from member ").append(IOwner); if (fVerbose) { sb.append(": ").append(parts); } } } } } } } return sb.toString(); } // ----- inner class: JMXPartitionStats ---------------------------------- /** * A class that calculate MBean attribute values from last sampled PartitionStats */ protected class JMXPartitionStats { /** * Return maximum node storage size. * * @return maximum node storage size in bytes */ public long getMaxStorageSize() { return m_cbMaxStorage; } /** * Return average node storage size. * * @return average node storage size in bytes */ public long getAverageStorageSize() { return m_cbAverageStorage; } /** * Return maximum partition storage size. * * @return Maximum partition storage size in bytes */ public long getMaxPartitionSize() { return m_cbMaxPartition; } /** * Return average partition storage size. * * @return average partition storage size in bytes */ public long getAveragePartitionSize() { return m_cbAveragePartition; } /** * Return node ID with maximum node storage size. * * @return node ID with maximum node storage size */ public int getMaxLoadNodeId() { return m_nMaxLoadId; } /** * Calculate MBean attribute values for partition statistics. */ public void calculateJMXPartitionStats() { DistributionManager manager = getManager(); if (manager == null) { // not initialized yet return; } PartitionStatistics[] aStats = manager.getPartitionStats(); Set setOwners = manager.getOwnershipMembers(); if (setOwners == null) { return; } long cbTotalStorage = 0; long cbMaxPartition = 0; long cbMaxStorage = 0; int nMaxNodeId = 0; for (Member member : setOwners) { PartitionSet parts = manager.getOwnedPartitions(member, 0); long cTotalStorageNode = 0; for (int iPart = parts.next(0); iPart >= 0; iPart = parts.next(iPart + 1)) { PartitionStatistics stat = aStats[iPart]; if (stat == null) { continue; } long cStoragePart = stat.getStorageSize(); cTotalStorageNode += cStoragePart; if (cStoragePart > cbMaxPartition) { cbMaxPartition = cStoragePart; } } cbTotalStorage += cTotalStorageNode; if (cTotalStorageNode > cbMaxStorage) { cbMaxStorage = cTotalStorageNode; nMaxNodeId = member.getId(); } } m_cbAveragePartition = cbTotalStorage / (getPartitionCount() * 1024); m_cbAverageStorage = cbTotalStorage / (setOwners.size() * 1024); m_cbMaxPartition = cbMaxPartition / 1024; m_cbMaxStorage = cbMaxStorage / 1024; m_nMaxLoadId = nMaxNodeId; } // ----- data members ------------------------------- /** * The maximum node storage size in bytes. */ protected long m_cbMaxStorage; /** * The average node storage size in bytes. */ protected long m_cbAverageStorage; /** * The maximum partition storage size in bytes. */ protected long m_cbMaxPartition; /** * The average partition storage size in bytes. */ protected long m_cbAveragePartition; /** * The node ID with maximum node storage size in bytes. */ protected int m_nMaxLoadId; } // ----- MBean support ------------------------------------------------ /** * Register an MBean representing this SimpleAssignmentStrategy. *

* This implementation creates and registers a {@link SimpleStrategyMBean}. */ protected void registerMBean() { PartitionedService service = getManager().getService(); Registry registry = service.getCluster().getManagement(); if (registry != null) { try { registry.register(registry.ensureGlobalName(makeMBeanName(service)), new AnnotatedStandardEmitterMBean(this, SimpleStrategyMBean.class)); } catch (NotCompliantMBeanException e) { throw Base.ensureRuntimeException(e); } } } /** * Create a name for the MBean representing this strategy. *

* The name must be globally unique, but not contain the nodeId property. * This allows us to re-bind the same JMX name to the MBean on a different node * when the distribution coordinator migrates. * * @param service partitioned service that uses this strategy * * @return the name for the MBean */ protected String makeMBeanName(PartitionedService service) { return Registry.PARTITION_ASSIGNMENT_TYPE + ",service=" + service.getInfo().getServiceName() + "," + Registry.KEY_RESPONSIBILITY + "DistributionCoordinator"; } /** * Unregister the MBean representing this SimpleAssignmentStrategy. */ protected void unregisterMBean() { PartitionedService service = getManager().getService(); Registry registry = service.getCluster().getManagement(); if (registry != null) { registry.unregister(registry.ensureGlobalName(makeMBeanName(service))); } } /** * Emit a partition loss notification for given partitions. * * @param partsLost the partition id that has been lost */ protected void emitLossNotification(PartitionSet partsLost) { PartitionedService service = getManager().getService(); Registry registry = service.getCluster().getManagement(); if (registry != null) { Member memberThis = service.getCluster().getLocalMember(); Notification note = new Notification(NOTIFY_LOST, String.valueOf(memberThis), -1, partsLost.cardinality() + " partitions have been lost"); note.setUserData(partsLost.toString()); registry.getNotificationManager().trigger( registry.ensureGlobalName(makeMBeanName(service)), note); } } /** * Collect the scheduled partition distributions, grouped by the primary owner * and storage index. A partition distribution (either primary or backup) is * considered to be scheduled if the distribution has been suggested but the * ownership change has not yet occurred. * * @return a Map of partition sets still awaiting distribution keyed by primary owner */ protected Map collectScheduledDistributions() { Map mapSuggestion = m_mapSuggestLast; if (mapSuggestion == null || mapSuggestion.isEmpty()) { return Collections.emptyMap(); } int cBackups = getBackupCount(); int cPartitions = getPartitionCount(); ServiceInfo serviceInfo = getManager().getService().getInfo(); Map mapScheduled = new HashMap<>(); for (Map.Entry entry : mapSuggestion.entrySet()) { Ownership ownersSuggest = entry.getKey(); PartitionSet partsSuggest = entry.getValue(); for (int iStore = 0; iStore <= cBackups; iStore++) { Member member = serviceInfo.getServiceMember(ownersSuggest.getOwner(iStore)); if (member != null) { PartitionSet partsScheduled = getUnownedPartitions(partsSuggest, member.getId()); if (!partsScheduled.isEmpty()) { PartitionSet[] aParts = mapScheduled.get(member); if (aParts == null) { mapScheduled.put(member, aParts = new PartitionSet[cBackups + 1]); } PartitionSet partsMember = aParts[iStore]; if (partsMember == null) { aParts[iStore] = partsMember = new PartitionSet(cPartitions); } partsMember.add(partsScheduled); } } } } return mapScheduled; } /** * Calculate a subset of the PartitionSet consisting of partitions that are * not owned (primary or backup) by the specified member. * * @param parts set of partitions to check the ownership of * @param nMember the member-id * * @return the subset of partitions that are not owned by the specified member */ protected PartitionSet getUnownedPartitions(PartitionSet parts, int nMember) { PartitionSet partsUnowned = new PartitionSet(parts); int cBackups = getBackupCount(); DistributionManager mgr = getManager(); for (int i = parts.next(0); i >= 0; i = parts.next(i + 1)) { Ownership owners = mgr.getPartitionOwnership(i); for (int iStore = 0; iStore <= cBackups; iStore++) { if (owners.getOwner(iStore) == nMember) { partsUnowned.remove(i); break; } } } return partsUnowned; } /** * Split the partition set scheduled for distribution by the current * primary owner (all transfers originate from the primary owner). * * @param parts a set of partitions not yet transferred * * @return the partitions scheduled for distribution, associated with their * primary owners */ protected SortedMap splitByOwner(PartitionSet parts) { SortedMap mapByOwner = new TreeMap<>(); int cPartitions = getPartitionCount(); DistributionManager mgr = getManager(); for (int nPid = parts.next(0); nPid >= 0; nPid = parts.next(nPid + 1)) { Ownership ownersCurrent = mgr.getPartitionOwnership(nPid); int nOwnerCurrent = ownersCurrent.getPrimaryOwner(); if (nOwnerCurrent == 0) { // orphaned partition; wait for the restore/recovery continue; } Integer IOwnerCurrent = Integer.valueOf(nOwnerCurrent); PartitionSet partsOwned = mapByOwner.get(IOwnerCurrent); if (partsOwned == null) { mapByOwner.put(IOwnerCurrent, partsOwned = new PartitionSet(cPartitions)); } partsOwned.add(nPid); } return mapByOwner; } // ----- inner interface: LoadCalculator ------------------------------ /** * LoadCalculator is used to calculate the scalar load (expressed as an * integer) of a partition (or set of partitions). */ public static interface LoadCalculator { /** * Return the load for the specified partition. * * @param iPartition the partition to determine the load for * * @return the load for the specified partition */ public int getLoad(int iPartition); /** * Return the load for the specified set of partitions. * * @param parts the partition set to determine the load for * * @return the load for the specified set of partitions */ public int getLoad(PartitionSet parts); } // ----- inner class: SimpleLoadCalculator ---------------------------- /** * Instantiate the load calculator. * * @param fPrimary true iff the load calculator will be used for primary * partition load; backup otherwise * * @return a load calculator */ public LoadCalculator instantiateLoadCalculator(boolean fPrimary) { return new SimpleLoadCalculator(); } /** * SimpleLoadCalculator defines a "count-based" load (e.g. the load of each * partition is defined to be 1). */ public static class SimpleLoadCalculator implements LoadCalculator { /** * {@inheritDoc} */ public int getLoad(int nPartition) { return 1; } /** * {@inheritDoc} */ public int getLoad(PartitionSet parts) { return parts.cardinality(); } } // ----- inner class: AnalysisContext --------------------------------- /** * Factory method. * * @return a new AnalysisContext */ public AnalysisContext instantiateAnalysisContext() { return new AnalysisContext(); } /** * AnalysisContext holds the working view of the partition ownership that is * used throughout the analysis and is used to reflect changes made during * this analysis. */ protected class AnalysisContext { // ----- constructors --------------------------------------------- /** * Default constructor. */ public AnalysisContext() { initialize(); } // ----- accessors ------------------------------------------------ /** * Return the set of updated partitions; may be null. * * @return the set of updated partitions, or null */ protected PartitionSet getUpdatedPartitions() { return m_partsUpdated; } /** * Return the BackupStrength for this analysis context. The backup * strength determines the degree of resiliency that the resulting * distribution will ensure (e.g. machine-safe, rack-safe, site-safe). * * @return the backup strength */ protected BackupStrength getBackupStrength() { return m_strength; } /** * Set the BackupStrength for this analysis context. * * @param strength the backup strength */ protected void setBackupStrength(BackupStrength strength) { m_strength = strength; } /** * Return the set of members across which to distribute the partition * ownership. *

* Note: The set of ownership members does not include any members that * may be in the process of leaving * * @return the set of (non-leaving) ownership enabled members */ protected Set getOwnershipMembers() { return m_setOwnershipMembers; } /** * Return the set of ownership members that are leaving. * * @return the set of leaving ownership enabled members */ protected Set getLeavingOwners() { return m_setOwenersLeaving; } /** * Return an array containing the members across which to distribute the * partition ownership, arranged in arbitrary order. *

* Note: The array does not include any members that may be in the * process of leaving * * @return an array containing the (non-leaving) ownership enabled members */ protected Member[] getOwnershipMembersList() { return m_aOwnershipMembers; } /** * Return the LoadCalculator used to calculate the primary partition load. * * @return the primary partition load calculator */ public LoadCalculator getPrimaryLoadCalculator() { return m_calculatorPrimary; } /** * Return the LoadCalculator used to calculate the backup partition load. * * @return the backup partition load calculator */ public LoadCalculator getBackupLoadCalculator() { return m_calculatorBackup; } /** * Return the number of backups to maintain, given the actual set of * ownership-enabled and leaving members. * * @return the number of backups to maintain */ protected int getActualBackupCount() { return m_cBackupActual; } /** * Return the time at which the analysis associated with this context * was completed, or 0 if it has not been completed. * * @return the time at which the analysis was completed, or 0 */ public long getCompletedTime() { return m_ldtCompleted; } /** * Set the timestamp at which the analysis associated with this context * completed. * * @param ldt the completion timestamp, or 0 */ protected void setCompletedTime(long ldt) { m_ldtCompleted = ldt; } /** * Return the partitions deemed orphaned as a result of a previous * execution of {@link PartitionAssignmentStrategy#analyzeOrphans(Map) * analyzeOrphans}. * * @return the partitions deemed orphaned after executing analyzeOrphans */ protected PartitionSet getOrphanedPartitions() { return m_partsOrphaned; } /** * Set the orphaned partitions that can be prioritized for transfer in * order to reduce the transfer cost. * * @param parts the set of orphaned partitions */ protected void setOrphanedPartitions(PartitionSet parts) { m_partsOrphaned = parts; } /** * Return the number of milliseconds the analysis should be delayed; * {@code 0L} suggests immediate analysis. * * @return the number of milliseconds the analysis should be delayed */ protected long getAnalysisDelay() { return m_cDelay; } /** * Set the number of milliseconds the analysis should be delayed; * {@code 0L} suggests immediate analysis. * * @param cDelay the number of milliseconds the analysis should be * delayed */ protected void setAnalysisDelay(long cDelay) { m_cDelay = cDelay; } /** * Reset those attributes that should be transient between {@link * #analyzeDistribution} requests. */ protected void resetTransients() { setOrphanedPartitions(null); setAnalysisDelay(-1); } // ----- AnalysisContext methods ---------------------------------- /** * Initialize the AnalysisContext. */ protected void initialize() { DistributionManager manager = getManager(); Set setOwners = manager.getOwnershipMembers(); Set setLeaving = manager.getOwnershipLeavingMembers(); if (!setLeaving.isEmpty()) { setOwners = new SubSet(setOwners); setOwners.removeAll(setLeaving); } // cache the set of non-leaving ownership members m_setOwnershipMembers = setOwners; m_aOwnershipMembers = (Member[]) setOwners.toArray(new Member[setOwners.size()]); m_setOwenersLeaving = setLeaving; // the number of backups could be smaller than configured if we have a // small enough number of members m_cBackupActual = Math.min(getBackupCount(), setOwners.size() - 1); // instantiate the level of backup strength m_strength = instantiateBackupStrength(setOwners); // instantiate the load calculators m_calculatorPrimary = instantiateLoadCalculator(true); m_calculatorBackup = instantiateLoadCalculator(false); // calculate and cache the fair-share load m_cFairSharePrimary = calculateFairShare(true); m_cFairShareBackup = calculateFairShare(false); } /** * Copy transient values from another, generally the previous, AnalysisContext * to this AnalysisContext. This provides an opportunity for the other * context to impart knowledge to this context. * * @param ctxLast the previous AnalysisContext */ protected void copyTransients(AnalysisContext ctxLast) { if (ctxLast == null) { return; } PartitionSet partsOrphaned = ctxLast.getOrphanedPartitions(); long cDelay = ctxLast.getAnalysisDelay(); if (partsOrphaned != null) { setOrphanedPartitions(partsOrphaned); } if (cDelay >= 0) { setAnalysisDelay(cDelay); } } /** * Return the (primary or backup) fair-share partition load. * * @param fPrimary true iff the primary fair-share should be returned * * @return the fair-share partition load */ protected int getFairShare(boolean fPrimary) { return fPrimary ? m_cFairSharePrimary : m_cFairShareBackup; } /** * Return the "fair share" (F) load. It is a ceiling for the load on a * for fully balance distribution. The fairness goal is to achieve such a * state that the load for all members is between F-L and F, where L is * typically the minimum of partition load values. * * @param fPrimary true for the primary fair-share, or false for backup * * @return the fair share */ protected int calculateFairShare(boolean fPrimary) { Set setOwners = getOwnershipMembers(); LoadCalculator calculator = fPrimary ? getPrimaryLoadCalculator() : getBackupLoadCalculator(); int cMembers = setOwners.size(); PartitionSet partsAll = new PartitionSet(getPartitionCount()); int cLoadTotal; partsAll.fill(); cLoadTotal = calculator.getLoad(partsAll); if (!fPrimary) { cLoadTotal *= getActualBackupCount(); } return cMembers <= 1 ? cLoadTotal : cLoadTotal/cMembers + 1; } /** * Return true iff the specified member is in the process of leaving. * * @param member the member * * @return true iff the specified member is in the process of leaving */ protected boolean isMemberLeaving(Member member) { return getLeavingOwners().contains(member); } /** * Create a backup strength to be used for distribution among the * specified set of ownership members. * * @param setOwners the ownership members * * @return the backup strength */ protected BackupStrength instantiateBackupStrength(Set setOwners) { // split the members by machine, rack, site Map mapBySite = new HashMap(); Map mapByRack = new HashMap(); Map mapByMachine = new HashMap(); for (Iterator iter = setOwners.iterator(); iter.hasNext(); ) { Member member = (Member) iter.next(); String sSite = member.getSiteName(); String sRack = member.getRackName(); int nMachine = member.getMachineId(); Set setSite = (Set) mapBySite.get(sSite); if (setSite == null) { setSite = new HashSet(); mapBySite.put(sSite, setSite); } setSite.add(member); Set setRack = (Set) mapByRack.get(sRack); if (setRack == null) { setRack = new HashSet(); mapByRack.put(sRack, setRack); } setRack.add(member); Set setMachine = (Set) mapByMachine.get(nMachine); if (setMachine == null) { setMachine = new HashSet(); mapByMachine.put(nMachine, setMachine); } setMachine.add(member); } // determine the highest level of safety that is achievable int nStrength = BackupStrength.NODE_SAFE; if (isStrongPossible(setOwners, mapBySite)) { nStrength = BackupStrength.SITE_SAFE; } else if (isStrongPossible(setOwners, mapByRack)) { nStrength = BackupStrength.RACK_SAFE; } else if (isStrongPossible(setOwners, mapByMachine)) { nStrength = BackupStrength.MACHINE_SAFE; } return new BackupStrength(nStrength, new HashSet(mapBySite.keySet()), new HashSet(mapByRack.keySet()), new HashSet(mapByMachine.keySet())); } /** * Return true iff a "strong" balanced distribution is achievable for the * specified set of members, split among a set of categories * (e.g. machine, rack, site). * * @param setOwners the set of ownership members * @param mapSplit the map of members, associated by their category * * @return true iff a "strong" balanced distribution is achievable */ protected boolean isStrongPossible(Set setOwners, Map mapSplit) { // Consider the number of members belonging to a specific group (e.g. a // machine, rack or site) to be given by m_i. If there are K distinct // groups the counts of members-per-group can be expressed as // m_1, m_2, ... m_K where: // // 1) let CMembers = Sum(m_1, ..., m_K) // 2) m_1 <= m_2 <= ... <= m_(K-1) <= m_K // // With a backup-count = B, the "worst-case" number of members' worth of // primary partitions that could be lost and should be tolerated is: // // sum of the members in the largest B groups // 3) let MaxLost = m_(K-B-1), ... m_(K-1), m_K // // In order for the loss of that many members' worth of partitions to be // tolerated, the surviving members must have enough backup-storage (in // the trivially balanced case) to compensate: // // 4) B (CMembers - MaxLost) >= MaxLost // // therefore, it follows: // // B*CMembers - B*MaxLost >= MaxLost // B*CMembers >= MaxLost + B*MaxLost // B*CMembers >= (B + 1)*MaxLost // // If we take a conservative approximation that: // // m_(K-B-1) ~= m_(K-B-2) ~= ... ~= m_(K-1) ~= m_K // // then MaxLost is conservatively estimated as: // // MaxLost ~= B*m_K // // therefore: // B*CMembers >= (B + 1)*(B*m_K) // B*CMembers >= B*(B + 1)*(m_K) // CMembers >= (B + 1)*(m_K) int cMax = 0; for (Iterator iter = mapSplit.values().iterator(); iter.hasNext(); ) { Set setGroup = (Set) iter.next(); int cMembers = setGroup.size(); if (cMembers > cMax) { cMax = cMembers; } } return cMax * (getBackupCount() + 1) <= setOwners.size(); } /** * Return true iff the ownership of the specified partition is "strong", * as defined by the current BackupStrength. * * @param iPartition the partition * * @return true iff the specified partition is strong */ protected boolean isPartitionStrong(int iPartition) { return isPartitionStrong(getPartitionOwnership(iPartition)); } /** * Return true iff the specified ownership is "strong", as defined by the * current BackupStrength. * * @param owners the ownership * * @return true iff the specified ownership is strong */ protected boolean isPartitionStrong(Ownership owners) { BackupStrength strength = getBackupStrength(); int cBackups = getActualBackupCount(); Member[] aMembers = new Member[cBackups + 1]; for (int iStore = 0; iStore <= cBackups; iStore++) { Member member = getMember(owners.getOwner(iStore)); if (member == null) { // orphaned/endangered cannot be "strong" return false; } aMembers[iStore] = member; } // compare the pair-wise strength among all owners for (int iStoreThis = 0; iStoreThis <= cBackups; iStoreThis++) { for (int iStoreThat = iStoreThis + 1; iStoreThat <= cBackups; iStoreThat++) { if (!strength.isStrong(aMembers[iStoreThis], aMembers[iStoreThat])) { return false; } } } return true; } /** * Return true iff the specified partition transfer would result in a * "strong" ownership, as defined by the current BackupStrength. * * @param iPartition the partition to transfer * @param iStore the storage index to transfer * @param member the member receiving the transfer * * @return true iff the specified partition transfer is strong */ protected boolean isTransferStrong(int iPartition, int iStore, Member member) { Ownership owners = (Ownership) getPartitionOwnership(iPartition).clone(); owners.setOwner(iStore, member.getId()); return isPartitionStrong(owners); } /** * Return true iff the specified members are mutually "strong", as * defined by the backup strength. * * @param member1 the first member to compare * @param member2 the second member to compare * * @return true iff the specified members are mutually strong */ protected boolean isStrong(Member member1, Member member2) { return getBackupStrength().isStrong(member1, member2); } /** * Return true iff the specified member is "strong" with respect to the * specified ownership, as defined by the backup strength. * * @param member the member * @param owners the ownership * * @return true iff the member is "strong" with respect to the ownership */ protected boolean isStrong(Member member, Ownership owners) { int cBackups = getBackupCount(); for (int iStore = 0; iStore <= cBackups; iStore++) { int nOwner = owners.getOwner(iStore); if (nOwner != 0 && isStrong(member, getMember(nOwner))) { return true; } } return false; } /** * Return a partition set representing the subset of the specified * partitions that are "orphaned". * * @param parts the partition set to collect * * @return a partition set containing the orphaned partitions */ protected PartitionSet collectOrphaned(PartitionSet parts) { PartitionSet partsOrphaned = getOrphanedPartitions(); PartitionSet partsOwnedOrphans; if (partsOrphaned != null && partsOrphaned.intersects(parts)) { partsOwnedOrphans = new PartitionSet(partsOrphaned); partsOwnedOrphans.retain(parts); } else { partsOwnedOrphans = new PartitionSet(parts.getPartitionCount()); } return partsOwnedOrphans; } /** * Return a partition set representing the subset of the specified * partitions that are "weak" or "vulnerable" (as defined by the backup * strength). * * @param parts the partition set to collect * * @return a partition set containing the weak partitions */ protected PartitionSet collectWeak(PartitionSet parts) { PartitionSet partsWeak = new PartitionSet(parts.getPartitionCount()); for (int iPart = parts.next(0); iPart >= 0; iPart = parts.next(iPart + 1)) { if (!isPartitionStrong(iPart)) { partsWeak.add(iPart); } } return partsWeak; } /** * Return true iff the specified partition is "endangered". A partition is * "endangered" if it is incompletely backed up (e.g. some backup copies * do not exist). * * @param iPartition the partition to test the endangered status * * @return true iff the specified partition is endangered */ protected boolean isPartitionEndangered(int iPartition) { return isPartitionEndangered(getPartitionOwnership(iPartition)); } /** * Return true iff the specified ownership is "endangered". A partition * is "endangered" if it is incompletely backed up (e.g. some backup * copies do not exist). * * @param owners the ownership to test for endangered status * * @return true iff the specified partition is endangered */ protected boolean isPartitionEndangered(Ownership owners) { int cBackups = getActualBackupCount(); for (int iStore = 1; iStore <= cBackups; iStore++) { if (owners.getOwner(iStore) == 0) { return true; } } return false; } /** * Return a partition set representing the subset of the specified * partitions that are "endangered". * * @param parts the partition set to collect * * @return a partition set containing the endangered partitions */ protected PartitionSet collectEndangered(PartitionSet parts) { PartitionSet partsEndangered = new PartitionSet(parts.getPartitionCount()); for (int iPart = parts.next(0); iPart >= 0; iPart = parts.next(iPart + 1)) { if (isPartitionEndangered(iPart)) { partsEndangered.add(iPart); } } return partsEndangered; } /** * Ensure and return the set of updated partitions. * * @return the set of updated partitions */ protected PartitionSet ensureUpdatedPartitions() { PartitionSet partsUpdated = getUpdatedPartitions(); if (partsUpdated == null) { m_partsUpdated = partsUpdated = new PartitionSet(getPartitionCount()); } return partsUpdated; } /** * Return the set of partitions for which the specified member owns (or * has been assigned by this analysis to own) the specified storage * index. * * @param member the member * @param iStore the storage index * * @return the set of partitions owned by the member at the specified index */ public PartitionSet getOwnedPartitions(Member member, int iStore) { PartitionSet[] aParts = m_mapOwnedPartitions.get(member); if (aParts == null) { aParts = new PartitionSet[1 + getBackupCount()]; m_mapOwnedPartitions.put(member, aParts); } PartitionSet parts = aParts[iStore]; if (parts == null) { parts = getManager().getOwnedPartitions(member, iStore); aParts[iStore] = parts; } return parts; } /** * Check if the distribution is in the initial state, when the coordinator * owns all the partitions and there are no backups. * * @param memberCoordinator the coordinator * * @return true if the coordinator owns all the partitions and there * are no backups */ public boolean isInitialDistribution(Member memberCoordinator) { PartitionedService service = getManager().getService(); int cBackups = service.getBackupCount(); for (int iPart = 0, c = service.getPartitionCount(); iPart < c; iPart++) { if (service.getPartitionOwner(iPart) != memberCoordinator) { return false; } for (int iBackup = 1; iBackup <= cBackups; iBackup++) { if (service.getBackupOwner(iPart, iBackup) != null) { return false; } } } return true; } /** * Set the context to pretend to be the "two servers" membership. * * @param member1 the first member * @param member2 the second member */ protected void primeDistribution(Member member1, Member member2) { int cFairInitial = getPartitionCount() / 2 + 1; Member[] aMembers = new Member[] {member1, member2}; m_setOwnershipMembers = new ImmutableArrayList(aMembers); m_aOwnershipMembers = aMembers; m_cFairSharePrimary = cFairInitial; m_cFairShareBackup = cFairInitial; } /** * Return the Ownership information (or the ownership assigned by this * analysis) for the specified partition. * * @param iPartition the partition to return the ownership for * * @return the Ownership information */ public Ownership getPartitionOwnership(int iPartition) { Ownership owners = m_aOwners[iPartition]; if (owners == null) { owners = getManager().getPartitionOwnership(iPartition); m_aOwners[iPartition] = owners; } return owners; } /** * Return the load (as defined by the appropriate load calculator) for * the specified partition. * * @param iPartition the partition to determine the load of * @param fPrimary true iff the primary load should be returned, or * false for backup * * @return the load for the specified partition */ protected int getPartitionLoad(int iPartition, boolean fPrimary) { return (fPrimary ? getPrimaryLoadCalculator() : getBackupLoadCalculator()) .getLoad(iPartition); } /** * Return the (primary or backup) partition load of the specified member. * * @param member the member to calculate the partition load for * @param fPrimary true for primary partition load, else backup load * * @return the partition load for the specified member */ protected int getMemberLoad(Member member, boolean fPrimary) { if (fPrimary) { return getPrimaryLoadCalculator().getLoad( getOwnedPartitions(member, 0)); } else { LoadCalculator calculator = getBackupLoadCalculator(); int cBackups = getBackupCount(); int cLoad = 0; for (int iStore = 1; iStore <= cBackups; iStore++) { cLoad += calculator.getLoad( getOwnedPartitions(member, iStore)); } return cLoad; } } /** * Calculate whether the analysis should be delayed. * * @return the delay before the next analysis, or 0 if the analysis * should commence immediately */ protected long calculateAnalysisDelay() { AnalysisContext ctxLast = getLastAnalysisContext(); if (ctxLast == null) { return 0; } // check to see what, if any, membership changes have occurred since // the last time analysis was considered Set setOwnersPrev = getLastOwnershipMembers(); Set setOwnersCur = getOwnershipMembers(); long cDelay = m_cDelay; if (cDelay >= 0) { // if the delay was explicitly specified we pass it on once m_cDelay = -1; return cDelay; } else if (setOwnersCur.equals(setOwnersPrev)) { // no membership change has occurred since the last delay calculation; // check to see if the previously issued advice has been enacted PartitionSet parts = ctxLast.getUpdatedPartitions(); if (parts == null || !ctxLast.getLeavingOwners().isEmpty()) { // 1. no changes were suggested last time - no harm in proceeding // 2. there were members leaving (graceful shutdown) therefore // re-execute a plan to ensure we get to balance asap and/or // re-issue any dropped advice due to restore during a plan update return 0L; } // account for all the rejected or dropped advice PartitionSet partsIgnored = getManager().getIgnoredAdvice(); if (partsIgnored != null) { // there is no reason to remember about a given advice // that was ignored parts.remove(partsIgnored); } for (int iPart = parts.next(0); iPart >= 0; iPart = parts.next(iPart + 1)) { Ownership ownersSuggested = ctxLast.getPartitionOwnership(iPart); Ownership ownersCurrent = getPartitionOwnership(iPart); if (!ownersCurrent.equals(ownersSuggested)) { // there are still suggestions from the last analysis // that are pending or in-flight and we should delay the // analysis to give more time for those suggestions to // take effect because the intermediate state // (partially implemented suggestions) may yield different // (conflicting) suggestions, leading to unnecessary // transfers, or unbalanced distributions (see COH-14898) cDelay = Math.max(0, ctxLast.getCompletedTime() + getSuggestionCompletionDelay() - Base.getSafeTimeMillis()); if (setOwnersCur.equals(ctxLast.getOwnershipMembers())) { // no membership changes have occurred, so there is no // real rush to re-analyze; proceed as scheduled; However // we don't want service waiting too long to call back return Math.min(cDelay, getSuggestionDelay()); } else { // there was a membership change from the last analysis time; // reschedule aggressively return Math.min(cDelay, 1000); } } } // all previous suggestions have been enacted, and there // is no reason to delay; proceed with the next analysis return 0L; } else if (setOwnersCur.containsAll(setOwnersPrev)) { // some members have joined; delay to let the membership settle return getMemberJoinDelay(); } else { // some members have left; we need to re-analyze immediately in // order to address the endangered partitions return 0L; } } /** * Update the analysis context to reflect the suggested transfer of the * specified number of primary partitions between the specified members. * * @param iPartition the partition id to transfer * @param iStore the storage index to transfer * @param memberFrom the member to transfer partitions from, or null if * the partition storage-index was endangered * @param memberTo the member to transfer partitions to, or null if * the partition storage index should be endangered */ protected void transitionPartition( int iPartition, int iStore, Member memberFrom, Member memberTo) { int nMemberTo = memberTo == null ? 0 : memberTo.getId(); int cBackups = getBackupCount(); Ownership owners = getPartitionOwnership(iPartition); // update the ownership info if (memberFrom != null) { getOwnedPartitions(memberFrom, iStore).remove(iPartition); } if (memberTo != null) { getOwnedPartitions(memberTo, iStore).add(iPartition); } for (int i = 0; i <= cBackups; i++) { if (i == iStore) // the storage index we are transferring { owners.setOwner(iStore, nMemberTo); // If we are doing a primary transfer, the old primary could // either become a new backup, or it could release the // storage. Demote the old primary to be a backup iff it // increases backup strength. // // See PartitionedService.assignPrimaryPartition() if (iStore == 0 && memberTo != null && memberFrom != null && !isMemberLeaving(memberFrom)) { for (int j = 1; j <= cBackups; j++) { Member memberBackup = getMember(owners.getOwner(j)); // Note: assignPrimaryPartition() is hard-coded to compare // for machine-safety if (memberTo.getMachineId() != memberFrom.getMachineId() && (memberBackup == null || memberTo.getMachineId() == memberBackup.getMachineId())) { getOwnedPartitions(memberFrom, j).add(iPartition); if (memberBackup != null) { getOwnedPartitions(memberBackup, j).remove(iPartition); } owners.setOwner(j, memberFrom.getId()); break; } } } } else if (nMemberTo != 0 && owners.getOwner(i) == nMemberTo) { // the new owner already owns a different storage-index of // this partition; set it to be endangered owners.setOwner(i, 0); getOwnedPartitions(memberTo, i).remove(iPartition); } } ensureUpdatedPartitions().add(iPartition); } /** * Suggest any distribution that may have been collected by this analysis * context to the DistributionManager. * * @return true iff a distribution was suggested */ protected boolean suggestDistribution() { PartitionSet partsUpdated = getUpdatedPartitions(); if (partsUpdated == null) { m_mapSuggestLast = Collections.emptyMap(); return false; } int cPartitions = getPartitionCount(); Map mapSuggest = new HashMap(); DistributionManager manager = getManager(); for (int iPart = partsUpdated.next(0); iPart >= 0; iPart = partsUpdated.next(iPart + 1)) { Ownership owners = getPartitionOwnership(iPart); PartitionSet parts = (PartitionSet) mapSuggest.get(owners); if (!owners.equals(manager.getPartitionOwnership(iPart))) { // there has actually been a change if (parts == null) { parts = new PartitionSet(cPartitions); mapSuggest.put(owners, parts); } parts.add(iPart); } } for (Iterator iter = mapSuggest.entrySet().iterator(); iter.hasNext(); ) { Entry entry = (Entry) iter.next(); Ownership owners = (Ownership) entry.getKey(); PartitionSet parts = (PartitionSet) entry.getValue(); if (!parts.isEmpty()) { manager.suggest(parts, owners); } } m_mapSuggestLast = mapSuggest; return true; } // ----- inner class: NotOwnedFilter ------------------------------ /** * Instantiate and return a NotOwnedFilter with the specified ownership. * * @param owners the ownership * * @return a NotOwnedFilter */ public Filter instantiateNotOwnedFilter(Ownership owners) { return new NotOwnedFilter(owners); } /** * NotOwnedFilter is a Filter implementation used to evaluate Member * objects, and selects members who are not represented in the reference * ownership object. */ public class NotOwnedFilter implements Filter { /** * Construct a NotOwnedFilter with the specified ownership. * * @param owners the ownership */ public NotOwnedFilter(Ownership owners) { m_owners = owners; } // ----- accessors -------------------------------------------- /** * Return the ownership used to evaluate member safety. * * @return the ownership */ public Ownership getOwnership() { return m_owners; } // ----- Filter methods --------------------------------------- /** * {@inheritDoc} */ public boolean evaluate(Object o) { Ownership owners = getOwnership(); int nMember = ((Member) o).getId(); int cBackups = getActualBackupCount(); for (int iStore = 0; iStore <= cBackups; iStore++) { if (owners.getOwner(iStore) == nMember) { return false; } } return true; } // ----- data members ----------------------------------------- /** * The ownership */ protected Ownership m_owners; } // ----- inner class: SafetyFilter -------------------------------- /** * Instantiate and return a SafetyFilter with the specified ownership. * * @param owners the ownership * @param iStore the storage index at which to evaluate members for safety * * @return a SafetyFilter */ public Filter instantiateSafetyFilter(Ownership owners, int iStore) { return new SafetyFilter(owners, iStore); } /** * SafetyFilter is a Filter implementation used to evaluate Member * objects, and selects members that are "strong" with respect to the * reference ownership, as defined by the backup-strength. */ public class SafetyFilter implements Filter { /** * Construct a SafetyFilter with the specified ownership. * * @param owners the ownership * @param iStore the storage index at which to evaluate members for safety */ public SafetyFilter(Ownership owners, int iStore) { m_owners = (Ownership) owners.clone(); m_iStore = iStore; } // ----- accessors -------------------------------------------- /** * Return the ownership used to evaluate member safety. * * @return the ownership */ public Ownership getOwnership() { return m_owners; } /** * Return the storage index at which to evaluate members for safety. * * @return the storage index at which to evaluate members for safety */ public int getStorageIndex() { return m_iStore; } // ----- Filter methods --------------------------------------- /** * {@inheritDoc} */ public boolean evaluate(Object o) { Ownership owners = getOwnership(); owners.setOwner(getStorageIndex(), ((Member) o).getId()); return isPartitionStrong(owners); } // ----- data members ----------------------------------------- /** * The ownership. */ protected Ownership m_owners; /** * The storage index at which to evaluate members for safety. */ protected int m_iStore; } // ----- inner class: UnderloadedFilter --------------------------- /** * Instantiate a filter that matches members with an over-load. * * @param fPrimary true for primary partition load * * @return a filter that matches members with an over-load */ public Filter instantiateOverloadedFilter(boolean fPrimary) { return new NotFilter(new UnderloadedFilter(fPrimary)); } /** * Instantiate a filter that matches members with an under-load. * * @param fPrimary true for primary partition load * * @return a filter that matches members with an under-load */ public Filter instantiateUnderloadedFilter(boolean fPrimary) { return new UnderloadedFilter(fPrimary); } /** * UnderloadedFilter is a Filter implementation that is used to evaluate * Member objects, and selects those whose partition load is * "underloaded" in comparison to the fair-share load. */ public class UnderloadedFilter implements Filter { /** * Construct an UnderloadedFilter. * * @param fPrimary true iff the filter should compare primary * partition load */ protected UnderloadedFilter(boolean fPrimary) { m_fPrimary = fPrimary; m_cFairShare = AnalysisContext.this.getFairShare(fPrimary); } // ----- accessors -------------------------------------------- /** * Return true iff this Filter compares the primary partition load. * * @return true if comparing the primary partition load */ public boolean isPrimary() { return m_fPrimary; } /** * Return the fair-share partition load. * * @return the fair-share partition load */ public int getFairShare() { return m_cFairShare; } // ----- Filter methods --------------------------------------- /** * {@inheritDoc} */ public boolean evaluate(Object o) { return getMemberLoad((Member) o, isPrimary()) < getFairShare(); } // ----- data members ----------------------------------------- /** * The cached fair-share load. */ protected int m_cFairShare; /** * True iff this filter represents primary partition load. */ protected boolean m_fPrimary; } // ----- inner class: LoadComparator ------------------------------ /** * Return a comparator for primary or backup partition load. * * @param fPrimary true for primary, or false for backup * * @return a comparator for primary or backup partition load */ public LoadComparator instantiateLoadComparator(boolean fPrimary) { return new LoadComparator(fPrimary); } /** * LoadComparator is a Comparator that can be used to compare two Member * objects based on their partition load (as defined by the * LoadCalculator). *

* A member is ordered "lower" than another member if it has a lower * member load (as determined by the LoadCalculator). Members with * equivalent loads are ordered equivalently. *

* Note: This comparator does not define an ordering that is "consistent * with equals". If used in a context requiring such a natural * ordering, it should be chained with comparator that does * provide a natural ordering. */ public class LoadComparator implements Comparator { /** * Construct a LoadComparator. * * @param fPrimary true if the comparator should use the primary load, * or false for backup load */ public LoadComparator(boolean fPrimary) { m_fPrimary = fPrimary; } // ----- accessors -------------------------------------------- /** * Return true iff the comparator should use the primary load. * * @return true iff the comparator should use the primary load */ public boolean isPrimary() { return m_fPrimary; } // ----- Comparator methods ----------------------------------- /** * {@inheritDoc} */ public int compare(Object o1, Object o2) { Member member1 = (Member) o1; Member member2 = (Member) o2; boolean fPrimary = isPrimary(); int cLoad1 = getMemberLoad(member1, fPrimary); int cLoad2 = getMemberLoad(member2, fPrimary); return cLoad1 - cLoad2; } // ----- data members ----------------------------------------- /** * Flag for primary or backup load comparison. */ protected boolean m_fPrimary; } // ----- inner class: StrengthComparator -------------------------- /** * Instantiate a StrengthComparator for the specified reference ownership. * * @param owners the ownership, from which to determine member strength * * @return a StrengthComparator */ public StrengthComparator instantiateStrengthComparator(Ownership owners) { return new StrengthComparator(owners); } /** * StrengthComparator is an Comparator that can be used to compare two * Member objects based on their "distance" from a given set of members * (as represented by an Ownership object). Member distances are * expressed with the granularity of "machine", "rack", and "site". *

* A member is ordered "lower" than another member if it has a larger * combined distance from the reference members (ownership) (i.e. it is * "stronger"). Members with equivalent distances are ordered * arbitrarily. *

* Note: This comparator does not define an ordering that is "consistent * with equals". If used in a context requiring such a natural * ordering, it should be chained with comparator that does * provide a natural ordering. */ public class StrengthComparator implements Comparator { /** * Construct a StrengthComparator for the specified reference ownership. * * @param owners the ownership, from which to determine member * strength */ public StrengthComparator(Ownership owners) { m_owners = owners; } // ----- accessors -------------------------------------------- /** * Return the ownership to use in comparing member strength. * * @return the ownership */ public Ownership getOwnership() { return m_owners; } // ----- Comparator methods ----------------------------------- /** * {@inheritDoc} */ public int compare(Object o1, Object o2) { // larger distances appear "first" return getDistance((Member) o2) - getDistance((Member) o1); } // ----- helpers ---------------------------------------------- /** * Return the "distance" of the specified member from the reference * ownership. The distance reflects granularities of "machine", * "rack" and "site", and in the case of multiple-backups, reflects * combined distances to each backup. * * @param member the member to return the distance for * * @return the "distance" */ protected int getDistance(Member member) { Ownership owners = getOwnership(); int cBackups = owners.getBackupCount(); int nDistance = 0; // Calculate a distance metric to be used in comparing // various members' strength with respect to the // single reference ownership. While the resulting // metric needs to provide a consistent comparison // across multiple members, it does not hold any // meaning outside of the given reference ownership. // // For example, the reference ownership could take one // of 3 "shapes": // // (?, 0, 0) - no owners at all // (?, a, 0) - some storage indices are unowned // (?, a, b) - all storage indices are owned // // The "distance" metric will return: // (?, 0, 0) - 0 (all members are equally "distant") // (?, a, 0) - dist(m, a)^2 // (?, a, b) - dist(m, a)^2 + dist(m, b)^2 for (int iStore = 0; iStore <= cBackups; iStore++) { int nOwner = owners.getOwner(iStore); if (nOwner != 0) { Member memberBackup = getMember(nOwner); if (member == null || memberBackup == null) { Logger.err(String.format("%s is null for %s", member == null ? "Slot in member target array" : "Backup owner", owners)); } int n = getDistance(member, getMember(nOwner)); nDistance += n * n; } } return nDistance; } /** * Return the "distance" between the specified members. * * @param member1 the first member * @param member2 the second member * * @return the "distance" between the specified members */ protected int getDistance(Member member1, Member member2) { if (!Base.equals(member1.getSiteName(), member2.getSiteName())) { return BackupStrength.SITE_SAFE; } if (!Base.equals(member1.getRackName(), member2.getRackName())) { return BackupStrength.RACK_SAFE; } if (member1.getMachineId() != member2.getMachineId()) { return BackupStrength.MACHINE_SAFE; } if (member1.getId() != member2.getId()) { return BackupStrength.NODE_SAFE; } Base.azzert(member1 == member2); return 0; } // ----- data members ----------------------------------------- /** * The ownership. */ protected Ownership m_owners; } // ----- default comparator --------------------------------------- /** * Instantiate a default member Comparator. The returned comparator * must define a strict total ordering over the set of Members. In * other words, no two distinct members may be {@link Comparator#compare * compared} to be equivalent. * * @return a Comparator that defines a strict total ordering of Members */ public Comparator instantiateDefaultComparator() { // MEMBERID_COMPARATOR is immutable return MEMBERID_COMPARATOR; } // ----- data members --------------------------------------------- /** * The primary LoadCalculator. */ protected LoadCalculator m_calculatorPrimary; /** * The backup LoadCalculator. */ protected LoadCalculator m_calculatorBackup; /** * The map of member ownership information for this analysis context. */ protected Map m_mapOwnedPartitions = new HashMap<>(); /** * The ownership array for this analysis context. */ protected Ownership[] m_aOwners = new Ownership[getPartitionCount()]; /** * The set of partitions that have been updated in this analysis context; * may be null. */ protected PartitionSet m_partsUpdated; /** * The set of partitions that were determined to be orphaned in the call * to {@link PartitionAssignmentStrategy#analyzeOrphans(Map)}. */ protected PartitionSet m_partsOrphaned; /** * The backup strength for the resiliency of the resulting distribution. */ protected BackupStrength m_strength; /** * The number of backup storages to maintain. *

* Note: this may differ from the configured backup count if there is an * inadequate number of ownership members to sustain the configured * backup count. */ protected int m_cBackupActual; /** * The fair-share primary partition load. */ protected int m_cFairSharePrimary; /** * The fair-share backup partition load. */ protected int m_cFairShareBackup; /** * The set of ownership members to include in the distribution. *

* Note: this set does not include members that are leaving */ protected Set m_setOwnershipMembers; /** * The set of ownership members that are leaving. */ protected Set m_setOwenersLeaving; /** * An array of the ownership members to include in the distribution, * arranged in arbitrary order. This array could be used for algorithms * performing in-place sorting of the members. *

* Note: this list does not include members that are leaving */ protected Member[] m_aOwnershipMembers; /** * The timestamp of when the analysis represented by this context was * completed, or 0 if it is not yet complete. */ protected long m_ldtCompleted; /** * An explicit delay to be used in favor of a determined delay in * {@link #calculateAnalysisDelay()}. */ protected long m_cDelay = -1L; } // ----- inner class: BackupStrength ---------------------------------- /** * BackupStrength represents a level of "strength" or "resiliency" between * the primary and backup owners of a partition. The BackupStrength is used * to determine which backup owners could serve as a "strong" backup for a * primary owner. */ protected static class BackupStrength { /** * Construct a BackupStrength of the specified strength. * * @param nStrength one of the BackupStrength.*_SAFE constants * @param setSites the site names * @param setRacks the rack names * @param setMachines the machine names */ protected BackupStrength(int nStrength, Set setSites, Set setRacks, Set setMachines) { m_nStrength = nStrength; m_setSites = setSites; m_setRacks = setRacks; m_setMachines = setMachines; } // ----- BackupStrength methods ------------------------------- /** * Return the next weakest BackupStrength. * * @return a BackupStrength that is immediately weaker than this */ protected BackupStrength getWeaker() { int nStrength = m_nStrength; if (nStrength == NODE_SAFE) { // NODE_SAFE is the weakest possible strength throw new IllegalStateException( "NODE_SAFE is the weakest BackupStrength"); } return new BackupStrength(nStrength - 1, m_setSites, m_setRacks, m_setMachines); } /** * Return true iff the specified members are mutually "strong". * * @param member1 the first member to compare * @param member2 the second member to compare * * @return true iff the specified members are mutually "strong" */ protected boolean isStrong(Member member1, Member member2) { switch (m_nStrength) { default: case NODE_SAFE: return member1.getId() != member2.getId(); case MACHINE_SAFE: return member1.getMachineId() != member2.getMachineId(); case RACK_SAFE: return !Base.equals(member1.getRackName(), member2.getRackName()); case SITE_SAFE: return !Base.equals(member1.getSiteName(), member2.getSiteName()); } } /** * Return the site count. * * @return the site count */ public int getSiteCount() { return m_setSites.size(); } /** * Return the rack count. * * @return the rack count */ public int getRackCount() { return m_setRacks.size(); } /** * Return the site count. * * @return the site count */ public int getMachineCount() { return m_setMachines.size(); } /** * Return a human-readable description string of this backup-strength. * * @return a human-readable description string of this backup-strength */ public String getDescription() { switch (m_nStrength) { default: return "ENDANGERED"; case NODE_SAFE: return "NODE-SAFE"; case MACHINE_SAFE: return "MACHINE-SAFE"; case RACK_SAFE: return "RACK-SAFE"; case SITE_SAFE: return "SITE-SAFE"; } } // ----- Object methods ------------------------------------------- /** * {@inheritDoc} */ public String toString() { return "BackupStrength{" + getDescription() + "}"; } // ----- constants and data members ------------------------------- /** * Node-safety (members are different). */ protected static final int NODE_SAFE = 1; /** * Machine-safety (members are on different machines). */ protected static final int MACHINE_SAFE = 2; /** * Rack-safety (members are on different racks). */ protected static final int RACK_SAFE = 3; /** * Site-safety (members are on different sites). */ protected static final int SITE_SAFE = 4; /** * The strength (one of the *_SAFE constants). */ protected int m_nStrength; /** * The set of site names. */ protected Set m_setSites; /** * The set of rack names. */ protected Set m_setRacks; /** * The set of machine names. */ protected Set m_setMachines; } // ----- constants and data members ----------------------------------- /** * Comparator used to provide arbitrary (equals-compatible) comparisons * between members. */ protected static final Comparator MEMBERID_COMPARATOR = (o1, o2) -> ((Member) o1).getId() - ((Member) o2).getId(); /** * The message returned by SimpleStrategyMBean when the distribution coordinator has not done its * first analysis yet. */ protected static final String MSG_NO_RESULT = "There are no distribution analysis results."; /** * The message returned by SimpleStrategyMBean when all suggested distributions have completed * and none are in-progress or scheduled. */ protected static final String MSG_NO_PENDING = "No distributions are currently scheduled for this service."; /** * The DistributionManager. */ protected DistributionManager m_manager; /** * The last analysis context. */ protected AnalysisContext m_ctxLast; /** * The Set of ownership-enabled members at the time of the last analysis. */ protected Set m_setOwnersLast; /** * The Map containing the last distribution suggested by this strategy. */ protected Map m_mapSuggestLast; /** * The JMXPartitionStats that hold the last updated jmx attributes. */ protected JMXPartitionStats m_statsPartition; /** * True if JMXPartitionStats should be updated. */ protected boolean m_fRefresh; /** * Trivial distribution is a concept introduced in SE One, where a trivial * two servers topology behaves in a special way - holding all the primary * partitions on the ownership senior and all the backups on the other * server. As soon as a third server is added, this functionality should * be turned off for good and a standard algorithm take place. */ private boolean m_fTrivialDistribution = Config.getBoolean("coherence.distribution.2server", false); /** * The amount of time in ms to delay the analysis after a * distribution suggestion has been made and before it is carried out. */ protected long m_cPlanCompletionDelay; }





© 2015 - 2024 Weber Informatics LLC | Privacy Policy